The Improvement Research Of Pedestrian Detector Based On Feature Fusion

The applications of Pedestrian Detection technology involving multiple fields, has a broad application prospects. The accuracy of the existing pedestrian detection algorithm is not good enough. The environmental requirements are too harsh. This paper further study of pedestrian detection in complex environments, and strive to achieve a breakthrough in the theory and practical application.This paper focus on the research work in the following two ways:(1) Constitute the framework of the current pedestrian detector were detailed analysis and research, introduced description of the existing pedestrian common features and classifier. On the basis of previous study and theory, for pedestrian detection algorithm does not take into account detection performance and the speed of detection problem, we propose a multi-stage feature fusion, this feature use a few simple features can be quickly calculated, such as partial summation, integral image, gradient maps, etc., to approximating some main features with a good detection performance but computational complex. The features relative to other features of a single color, Haar-like, HOG,CSS, which are more abundant expression and can quickly calculate. Experimental data show that, the performance of improved algorithm based on feature fusion can be ensured and the speed of detection is faster significantly.(2) For classical HOG pedestrian detection, because the problem of waste time in the multi-scale traverse stages to establish the traditional image pyramid, we propose an improved algorithm based on scale approximating. There is similarity between the pyramid images at different scales, through the derivation of the gradient histogram, approximating the original HOG features to create a HOG features pyramids. Thus changed the traditional tedious steps to be seized from image to image pyramid space, and then calculate the feature space. Experimental data show that,based on the approximate scale can shorten time to traverse the image, effectively reducing the detection time, easy to achieve real-time pedestrian detection.